Fix #3485, #3540: Don't use dropout for predicting test sets (#3556)

* Fix #3485, #3540: Don't use dropout for predicting test sets

Dropout (for DART) should only be used at training time.

* Add regression test

include/xgboost/gbm.h

@@ -76,11 +76,14 @@ class GradientBooster {
    * \brief generate predictions for given feature matrix
    * \param dmat feature matrix
    * \param out_preds output vector to hold the predictions
+   * \param dropout whether dropout should be applied to prediction
+   *   This option is only meaningful if booster='dart'; otherwise ignored.
    * \param ntree_limit limit the number of trees used in prediction, when it equals 0, this means
    *    we do not limit number of trees, this parameter is only valid for gbtree, but not for gblinear
    */
   virtual void PredictBatch(DMatrix* dmat,
                             HostDeviceVector<bst_float>* out_preds,
+                            bool dropout = true,
                             unsigned ntree_limit = 0) = 0;
   /*!
    * \brief online prediction function, predict score for one instance at a time

src/gbm/gblinear.cc

@@ -103,6 +103,7 @@ class GBLinear : public GradientBooster {
 
   void PredictBatch(DMatrix *p_fmat,
                     HostDeviceVector<bst_float> *out_preds,
+                    bool dropout,
                     unsigned ntree_limit) override {
     monitor_.Start("PredictBatch");
     CHECK_EQ(ntree_limit, 0U)

src/gbm/gbtree.cc

@@ -217,6 +217,7 @@ class GBTree : public GradientBooster {
 
   void PredictBatch(DMatrix* p_fmat,
                HostDeviceVector<bst_float>* out_preds,
+               bool dropout,
                unsigned ntree_limit) override {
     predictor_->PredictBatch(p_fmat, out_preds, model_, 0, ntree_limit);
   }
@@ -356,8 +357,11 @@ class Dart : public GBTree {
   // predict the leaf scores with dropout if ntree_limit = 0
   void PredictBatch(DMatrix* p_fmat,
                     HostDeviceVector<bst_float>* out_preds,
+                    bool dropout,
                     unsigned ntree_limit) override {
-    DropTrees(ntree_limit);
+    if (dropout) {
+      DropTrees(ntree_limit);
+    }
     PredLoopInternal<Dart>(p_fmat, &out_preds->HostVector(), 0, ntree_limit, true);
   }
 

src/learner.cc

@@ -469,7 +469,7 @@ class LearnerImpl : public Learner {
     } else if (pred_leaf) {
       gbm_->PredictLeaf(data, &out_preds->HostVector(), ntree_limit);
     } else {
-      this->PredictRaw(data, out_preds, ntree_limit);
+      this->PredictRaw(data, out_preds, false, ntree_limit);
       if (!output_margin) {
         obj_->PredTransform(out_preds);
       }
@@ -560,14 +560,16 @@ class LearnerImpl : public Learner {
    * \brief get un-transformed prediction
    * \param data training data matrix
    * \param out_preds output vector that stores the prediction
+   * \param dropout whether dropout should be applied to prediction.
+   *   This option is only meaningful if booster='dart'; otherwise ignored.
    * \param ntree_limit limit number of trees used for boosted tree
    *   predictor, when it equals 0, this means we are using all the trees
    */
   inline void PredictRaw(DMatrix* data, HostDeviceVector<bst_float>* out_preds,
-                         unsigned ntree_limit = 0) const {
+                         bool dropout = true, unsigned ntree_limit = 0) const {
     CHECK(gbm_ != nullptr)
         << "Predict must happen after Load or InitModel";
-    gbm_->PredictBatch(data, out_preds, ntree_limit);
+    gbm_->PredictBatch(data, out_preds, dropout, ntree_limit);
   }
 
   // model parameter

tests/python/test_basic_models.py

@@ -48,6 +48,13 @@ class TestModels(unittest.TestCase):
         preds2 = bst2.predict(dtest2, ntree_limit=num_round)
         # assert they are the same
         assert np.sum(np.abs(preds2 - preds)) == 0
+        # regression test for issues #3485, #3540
+        for _ in range(10):
+            bst3 = xgb.Booster(params=param, model_file='xgb.model.dart')
+            dtest3 = xgb.DMatrix('dtest.buffer')
+            preds3 = bst3.predict(dtest3)
+            # assert they are the same
+            assert np.sum(np.abs(preds3 - preds)) == 0, 'preds3 = {}, preds = {}'.format(preds3, preds)
 
         # check whether sample_type and normalize_type work
         num_round = 50